How AI Reads Your Text: Tokens, Costs, and Context Windows Explained

When you send a prompt to ChatGPT, Claude, or Gemini, the model never sees your words the way you wrote them. Before anything happens, your text is chopped into small pieces called tokens — and almost everything that matters about working with AI, from the bill you pay to whether your prompt is accepted at all, is measured in those tokens. Once you understand them, a lot of confusing AI behaviour suddenly makes sense.

What exactly is a token?

A token is a chunk of text the model treats as a single unit. It is usually not a whole word and not a single letter — it sits somewhere in between. Common short words like "the" or "and" are one token each. Longer or rarer words get split into pieces: "tokenization" might become "token" + "ization". Spaces, punctuation, and line breaks count too.

That "depends on what you write" part is important. Code, with its brackets and symbols, tokenizes far less efficiently than prose. Emoji can cost several tokens each. And languages that do not use the Latin alphabet — Chinese, Japanese, Arabic — often use noticeably more tokens per character than English, because the tokenizer was trained mostly on English-heavy data.

Why tokens decide your bill

Every major AI API charges by the token, and it charges twice: once for the tokens you send (the input, or prompt) and again for the tokens the model generates back (the output, or completion). Output tokens are almost always more expensive than input tokens — often three to five times more — because generating text is the costly part.

Providers publish prices per million tokens, which can make small numbers look harmless. But costs scale with usage. A prompt that costs a fraction of a cent once becomes a real number when you run it across thousands of documents, or stuff a long history into every request. This is exactly why estimating before you build is worth the thirty seconds it takes.

The context window: will it even fit?

Every model has a context window — the maximum number of tokens it can hold in mind at once. Crucially, that budget covers both your input and the reply. If your prompt is 120,000 tokens and the window is 128,000, you have only 8,000 tokens left for the answer. Ask for more and the request fails or, worse, quietly truncates and gives you a half-formed response.

• GPT-4o family: 128,000 tokens
• Claude 3 family: 200,000 tokens
• Gemini 1.5 Pro: up to 2,000,000 tokens
• GPT-3.5 Turbo: 16,385 tokens — small, and easy to overflow with long documents

A bigger window is not automatically better. Larger context often costs more, and stuffing a window full of marginally-relevant text can actually make answers worse. The practical move is to check that your input plus a realistic reply comfortably fits — with headroom to spare.

Why token counts sometimes disagree

Paste the same sentence into two different counters and you may get two different numbers. That is not a bug — different model families use different tokenizers. OpenAI publishes its tokenizers openly, so a good counter can give you an exact GPT-4o or GPT-4 count. Anthropic and Google do not ship a public browser tokenizer, so any count for Claude or Gemini is a close approximation, not a guarantee.

Practical habits that save tokens (and money)

1. Trim the obvious fat — remove boilerplate, repeated instructions, and dead context from your prompts
2. Reserve only the output length you actually need; do not pay for a 4,000-token reply when 400 will do
3. Summarise long histories instead of resending the full transcript every turn
4. Pick the smallest model that does the job — GPT-4o mini and Claude Haiku are a fraction of the flagship price
5. Count before you scale: estimate one request, then multiply by your real volume

None of this requires being an AI expert. It just requires being able to see the numbers — and that is exactly the gap a good token counter fills. Paste your prompt, read the count, the cost, and the fit, and you are making an informed decision instead of a hopeful guess.